|
Accurate prediction of the surrounding traffic environment is crucial for the safety of autonomous vehicles. However, the limitation of onboard system resources and the complexity and diversity of driving scenes hinder the deployment of scene understanding in the auto-drive system. This paper optimizes the backbone network, uses deep separable convolution to reduce the complexity of network operations and uses multiple attention mechanisms in the decoding stage. On this basis, this paper adopts the shared strategy for the feature extraction module and jointly trains the semantic segmentation and Object detection, which can reduce the network parameters, improve the reasoning speed, and improve the accuracy. We have evaluated the proposed method on the public data set. The results show that our method achieves the most advanced performance and can balance speed and accuracy.
|
